Scikit F-score metric error

Question

I am trying to predict a set of labels using Logistic Regression from SciKit. My data is really imbalanced (there are many more '0' than '1' labels) so I have to use the F1 score metric during the cross-validation step to "balance" the result.

[Input]
X_training, y_training, X_test, y_test = generate_datasets(df_X, df_y, 0.6)
logistic = LogisticRegressionCV(
    Cs=50,
    cv=4,
    penalty='l2', 
    fit_intercept=True,
    scoring='f1'
)
logistic.fit(X_training, y_training)
print('Predicted: %s' % str(logistic.predict(X_test)))
print('F1-score: %f'% f1_score(y_test, logistic.predict(X_test)))
print('Accuracy score: %f'% logistic.score(X_test, y_test))

[Output]
>> Predicted: [0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0]
>> Actual:    [0 0 0 1 0 0 0 0 0 1 1 0 0 1 0 0 0 0 0 0 0 1 1]
>> F1-score: 0.285714
>> Accuracy score: 0.782609
>> C:\Anaconda3\lib\site-packages\sklearn\metrics\classification.py:958:  
   UndefinedMetricWarning:
   F-score is ill-defined and being set to 0.0 due to no predicted samples.

I certainly know that the problem is related to my dataset: it is too small (it is only a sample of the real one). However, can anybody explain the meaning of the "UndefinedMetricWarning" warning that I am seeing? What is actually happening behind the curtains?

Answer 1

It seems it is a known bug here which has been fixed, I guess you should try update sklearn.

Answer 2

However, can anybody explain the meaning of the "UndefinedMetricWarning" warning that I am seeing? What is actually happening behind the curtains?

This is well-described at https://stackoverflow.com/a/34758800/1587329:

https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/metrics/classification.py

F1 = 2 * (precision * recall) / (precision + recall)

precision = TP/(TP+FP) as you've just said if predictor doesn't predicts positive class at all - precision is 0.

recall = TP/(TP+FN), in case if predictor doesn't predict positive class - TP is 0 - recall is 0.

So now you are dividing 0/0.

To fix the weighting problem (it's easy for the classifier to (almost) always predict the more prevalent class), you can use class_weight="balanced":

logistic = LogisticRegressionCV(
    Cs=50,
    cv=4,
    penalty='l2', 
    fit_intercept=True,
    scoring='f1',
    class_weight="balanced"
)

LogisticRegressionCV says:

The “balanced” mode uses the values of y to automatically adjust weights inversely proportional to class frequencies in the input data as n_samples / (n_classes * np.bincount(y)).